Improving Retrieval Augmented Generation using LLMs

In a business context, specifically in the insurance sector, conversational RAG models grounded in a company's private information can have significant productivity impacts. Despite the potential of these systems, the high risk associated with the insurance sector makes the assertiveness of the system a crucial factor. Additionally, the existence of many different methods and distinct architectures — which must be chosen according to the specific case under study — makes evaluating these systems challenging. This dissertation aims to investigate recent RAG system techniques as well as the metrics used to evaluate them, with the goal of finding the best architecture for a system in the specific Fidelidade case study.

Techniques to explore

• Chunking, Indexing and Retrieval techniques for tabular data
• Chunking, Indexing and Retrieval techniques for images
• Agent frameworks with Reasoning for complex/iterative questions
• Portuguese-European Large Language Models

Cooperation with company or external entity: Fidelidade S.A.